Evaluation of Atlas Selection: How Close Are We to Optimal Selection?

Mark J. Gooding

CONTENTS

  • 3.1 Motivation for Atlas Selection................................................................................................19
  • 3.2 Methods of Atlas Selection.....................................................................................................22
  • 3.3 Evaluation of Image-Based Atlas Selection...........................................................................29
  • 3.3.1 Implementation...........................................................................................................29
  • 3.3.2 Brute-Force Search.....................................................................................................29
  • 3.3.3 Atlas Selection Performance Assessment...................................................................30
  • 3.3.4 Discussion and Implications for Atlas Selection........................................................32
  • 3.3.5 Limitations..................................................................................................................35
  • 3.4 Impact of Atlas Selection on Clinical Practice.......................................................................35
  • 3.5 Summary and Recommendations for Future Research..........................................................35

References........................................................................................................................................36

Image-based atlas selection has been used for more than two decades as an approach to improve atlas-based auto-contouring. Most published investigations have shown some level of improvement in the performance of auto-contouring over the random selection of atlases when using image-based selection. In this chapter, the published research into atlas-selection is reviewed, asking the question, “How close are we to optimal atlas selection?” An experiment is presented that assesses the most common approach to atlas selection - ranking similarity based on normalized mutual information (NMI) between the atlas and the test case - using the 2017 AAPM Thoracic Auto-segmentation Challenge data. In phrasing the evaluation with respect to optimality, it is seen that while there is some improvement in auto-contouring performance, atlas selection is far from optimal.

Motivation for Atlas Selection

The primary assumption of atlas-based contouring is that the anatomy of the person represented in the atlas is the same as that represented by the patient case to be contoured. While the majority of people are broadly anatomically similar, there is substantial diversity in size between subjects. Furthermore, there will be differences in position between patients when being imaged, leading to differences in image appearance. Therefore, deformable image registration is used to account for differences in size and positioning between subjects. However, deformable image registration is not perfect, and it is generally accepted that the less deformation that is required, the better the correspondence between images will be. Thus, it has been proposed that selecting an atlas that is more similar in appearance to the patient, thereby minimizing the deformation required, will result in a lower registration error and better atlas-segmentation performance than using an atlas that differs substantially from the patient in position or anatomy.

Exploring this in more depth, it is understood that deformable image registration is being used to account for differences in size and positioning between subjects. Yet, deformable image registration itself has no knowledge of anatomy and the goal of image registration is to map one image to another such that the appearance is the same. If taking this objective to the extreme, the optimum image similarity can be achieved by intensity sorting; the intensities of voxels in each image are put into ordered lists. Corresponding voxels would then be determined by the position in the ordered list, known as the “Completely Useless Registration Algorithm” [1]. The moving image can thus be deformed to match the fixed image, yet the deformation field and deformed image resulting from such a mapping are anatomically nonsense. Thus, deformable registration must be constrained by a regularizer to prevent non-meaningful correspondence. Neighboring voxels are expected to move in a similar way. However, this same regularizer can prevent fine scale alignment. So, while it is true that at a large scale the majority of patients (of the same sex) are anatomically similar, e.g. all patients will have a heart, and for most patients the heart lies to the left, at finer scales there are anatomical variations that may not adequately be addressed by deformable registration.

By way of example, take the degree to which the heart varies in position from patient to patient towards the left, as illustrated in Figure 3.1 for cases Train-S3-011 and Test-S3-102 from the thoracic challenge. Firstly, to anatomically match the heart from the training case, Train-S3-011, which could be used as an atlas to the test case, Test-S3-102, the heart must be moved further to the left. However, in the training case there is lung tissue between the heart and the ribcage that is not present in that location in the test case. Whether anatomically correct or not, this lung tissue must be pushed cranially to be out of the way. However, the organs below these regions of lung must not

Cases Train-S3-011

FIGURE 3.1 Cases Train-S3-011 (left) and Test-S3-102 (right) from the LSTSC. While the patients are anatomically the same, the position of the anatomy varies from subject to subject. Note particularly the difference in location, and orientation of the heart.

move cranially, nor the ribs towards the right. This anatomical difference between the patients cannot be overcome by a constrained registration. This is demonstrated in Figure 3.2 where the contours from Train-S3-011 have been propagated to Test-S3-102 following deformable registration. The left lung of the test case is reasonably well segmented on the coronal image shown using this particular image registration. Yet, to do so the regularizer of the deformable registration has dragged the caudal heart contour cranially, leading to a poor segmentation. In the axial view, the regularizer has prevented the heart moving left, leading to the lung contour going through the heart. While sliding boundary approaches have been proposed to allow for some relative organ motion, these still assume a one-to-one anatomical correspondence that may not be appropriate for intersubject registration.

Since deformable image registration struggles to overcome significant anatomical differences, it can be expected that if a more anatomically similar atlas is used then the registration, and consequently the contouring, would be better. Taking an extreme example again, if the images/anatomy are identical then there is no registration to do and the contouring should be perfect. In practice, it is observed that re-contouring using a previous time point of the same patient in a different position results in better auto-contouring than using an atlas based on a different person. This is illustrated in Figure 3.3, where re-contouring and atlas-based contouring are shown. In this figure the anterior aspect of the heart and the posterior boundary of the right lung show degradation in contour quality for atlas contouring (bottom) but are well contoured when using the same patient (top). Therefore,

Failure of atlas contouring resulting from anatomical variations

FIGURE 3.2 Failure of atlas contouring resulting from anatomical variations. LSTSC-Test-S3-102 (right) has been contoured following deformable registration with LSTSC-Train-S3-011 (right). Note the incorrect contouring of the heart.

Comparison of re-contouring with the same patient

FIGURE 3.3 Comparison of re-contouring with the same patient (top) and atlas contouring with a different subject (bottom). In each case the contours on the left have been mapped to the image on the right using deformable image registration. Note, the performance of atlas contouring at the posterior of the lung and left anterior of the heart is worse compared to re-contouring.

atlas selection has been proposed as a method to improve atlas-based auto-contouring, whereby an atlas (or atlases) is selected from a larger pool of available atlases with the intent that the selected atlas(es) is more anatomically similar to the patient.

Following this line of thought, if a more similar atlas will result in better contouring performance, then the larger the size of the database of atlases, the higher the expected performance for atlas contouring will be. Thus, atlas contouring may be sufficient to “solve” auto-contouring in radiotherapy if a large enough pool of atlases is created. This line of thinking was explored by Schipaanboord et al. [2], who used extreme value theory to predict the performance of atlas-based contouring in radiotherapy assuming a database size of 5000 atlases in the presence of perfect atlas selection. They concluded that such a database could yield performance equivalent to clinical variation.

 
Source
< Prev   CONTENTS   Source   Next >